Liquid AI Releases LFM2.5-2.6B: An On-Device Agentic Model With 128K Context, Tool Calling, And Open Weights

Liquid AI just put out LFM2.5-2.6B, and the pitch is almost boring in how practical it is: a 2.69-billion-parameter model that plans, calls tools, and chews through multi-step tasks entirely on the device in your pocket. No API call, no server round-trip, no per-token bill. Just a model sitting on your phone doing the work.

The specs are more interesting than the headline suggests. It packs a 131,072-token context window and a freshly doubled 128,000-token vocabulary, achieved by extending the existing tokenizer rather than starting over. Training ran on roughly 34 trillion tokens across 30 layers, mixing 22 gated short-convolution blocks with 8 grouped-query attention blocks — a hybrid architecture built for speed rather than raw scale. Liquid AI shipped two versions, a bare base checkpoint for people who want to fine-tune and a post-trained agentic version ready to run tasks out of the box, both under the lfm1.0 open license on Hugging Face.

Getting it running is where things get genuinely convenient. The weights come in native, GGUF, MLX, and ONNX formats, with support already baked into llama.cpp, vLLM, SGLang, and LM Studio on day one. On an M5 Max it decodes at 220 tokens per second while using under 2.5GB of memory. A single H100 can serve around 1.3 billion tokens a day if you'd rather self-host at scale. Fine-tuning via LoRA works with TRL and Unsloth, so a small team can realistically customize this without a research budget.

The post-training pipeline is the real story of why it performs the way it does. Liquid AI ran four stages: two rounds of supervised fine-tuning with a dataset seven times larger than what they used for their earlier 8B model, then domain-specific teacher specialization trained with reinforcement learning, then multi-domain distillation where the student learns under its own rollouts, and finally agentic reinforcement learning inside real tool-use environments like Hermes Agent and OpenClaw. That's a lot of machinery for a 2.6B model, and it shows in the benchmarks — it beats Qwen3.5-9B and Gemma-4-E4B-it on instruction-following tests like IFBench and Multi-IF, and edges most tool-calling benchmarks too, losing only on BFCLv4.

Where it doesn't compete is coding and deep knowledge tasks, and Liquid AI is upfront about that, scoring nearly 10 points behind Qwen3.5-9B on LiveCodeBench. This isn't trying to be a general-purpose brain. It's aimed squarely at regulated, offline, or air-gapped environments — automotive, healthcare, defense, industrial robotics — where keeping data on-device isn't a preference but a requirement.